The understanding of code smells have exerted profound inuence in the quality and the performance of programming codes. There are various type of code smells require various solutions. In order to interpret the solutions available in code smells, this research uses NLP (natural language programming) techniques to comprehend contents of messages from Technical Debt Dataset. Based on phrase structure rules, semantic patterns were extracted from the Dataset to build connection between trigger words and dependency tree. Verb Phrases are considered as the actions taken by programmers encountering code smells.
With the development of software design, the improvement of existing codes
have began to attract more and more attentions nowadays. In the perspective
of technical debt which re ect limited performance and lower e ciency of code,
there is an opportunity for programmers to interpret and update the existing
codes by code smells. In computer programming, codes smells were introduced
by Martin Fowler on improving the design of existing codes [
In this subsection, the method for extracting semantic patterns of the COMMIT CHANGE messages is discussed. The dynamics of extracting key perspectives includes focus; technique; domain. Then we describe about the data collection in the code smells domain. 2.1
Regarding code smells, there is a growing tendency of researches to evaluate
the quality of codes. NLP (natural language programming) tools can be used
to extract semantic patterns to identify the content of codes smells. Bird et al.
(2009) suggested the pipeline architecture for an information extraction system
(Fig. 1.) to extract structure data from raw text [
The messages of COMMIT CHANGES can explore the content of code smells with in-depth understanding of the fault proneness and valuable responding strategies. Encountering various code smells, COMMIT CHANGES messages are used for programmers to communicate towards various problems. In order to extract the contents of code smells, we use Phrase Structure Rules to extract semantic patterns as syntactic categories. For example, "You merge the branch in the trunk", Fig. 2. provides the sorting of reservation to claim the languagespeci c contexts.
In a certain type of research work, Gupta et al. provided a method for
characterizing into three aspects as Focus, Domain, and Techniques[
In the attempt to understand how the semantic patterns are composed in
the messages extracted from COMMIT CHANGES, Feature Request Mining
[
In the initial screening of the dataset, the commitHash tag represent a single COMMIT object, and the commitHash might be duplicated with the same messages to the same COMMIT, therefore, the duplicates of messages are removed based on commitHash. The cleaning process of the dataset provides one unique messages from each COMMIT object. 2.3
At beginning, all the raw texts with tags of COMMIT CHANGES are extracted. After initial screening, the tag of messages are extracted from the raw texts. After duplicate checking, there are 6307 unique messages. Because the research interests are in semantic patterns, the number, stopword, punctuation, plain text, white space are removed before tokenization. Each tokens are labelled by postagged methods to categorize into present tense verb (VBP), noun (NN), preposition (PRP), coordinating conjunction (CC), adjective (JJ) and son on. In order to extract most frequent used verbs, base (VB), 3rd singular present (VBZ), past participle (VBN), gerund (VBG), simple past (VBD) categories are collected to build the keywords for expressing feature requests.
In the next stage, the tag of Verb Phrase (VP in Fig. 2.) was extracted from the sentence, from T trigger word to d which is the dependency connecting with the trigger words, such as direct-object or subject. In feature request mining, it is the <request> or <existing feature>'s contents which represent the phrase structure rules. Such as, "use generic method for emptyMap instead of EMPTYMAP constant" can be clari ed into: use <request> instead of <existing feature>. 3
This section is to report the results and ndings of the methodologies applied in study design. The rst subsection illustrates verbs as keywords to express feature requests. The second subsection indicates how the phrase structure rules applied to extract <request> and <existing feature> from the sentences. 3.1
In order to extract the lexical categories of verbs, the NLTK python packages
were used to classify words into POS-tagging which tags words automatically
[
There are various constituent parts to form phrase structure rules. In this research, Verb Phrase (VP in Fig.2.) gives an internal structure of semantic patterns. A verbal phrase might consist V, NP, PP. In order to built the dependency remov, make, get, build, exist, failur, depend, refer, tri, expect, nd, origin, resourc, exclud, take, tracer, expand, know, occur, transit, accumuloconfdir, appear, gitignor, walk, detect, intend, think, span, fri, seem, lenam, mdrobclouderacom, suspend, favor, protect, remain, restor, deepcopi, sensit, accumuloenvsh, save, collect, hold, hope, startsh, accumuloinputformat, david, defend, prefer, referenc graph which is the Verb Phrase for semantic patterns, we use below Bi-gram formula to test the combinations of verb plus <request> or <existing feature> to form the connection between T trigger word to d direct-object or subject. Bigram uses the co-occurrence of two words, to divide the probability of the rst words. P() represents the conditional probability of the token and next one: W k; Wk 1: Bi-gram formula for probability on two adjacent elements on tokens.
P (Wk 1jWk) =
P (Wk 1; Wk)
P (Wk 1) (1)
After the calculation of Bi-gram, the semantic patterns were extracted based on the Table 2.'s verbs collection. For example, in Table 3. it is the "remove" verb based patterns extracted from the dataset. It can be seen clearly that "remove stray" takes 68.18 in Bi-gram result, and "remove references" takes 46.49. Based on the phrase structure rules, irrelevant semantic patterns were removed such as "remove the" and so on. In order to improve the quality and performance of codes, this research uses NLP techniques to interpret contents of messages in Technical Debt Dataset. Phrase structure rules are applied to extract semantic patterns from the dataset to understand techniques and methods used in the domain of codes smells. This research sheds the lights on addressing the contents of code smells by feature request mining. Moreover, this research can help programmers to understand the contents of code smells and what actions can be taken to improve the quality of codes. 4.1
The semantic pattern extraction is based on the phrase structure rules, with various phrase structure rules, the interpretation of the results may vary a lot. This study apples the typical of V + NP, other rules can also be applied, such as NP = Det (determiner) + N (noun) and so on. Deeply, based on the domain of code smells, speci c trigger words and dependency tree can be collected to understand the contents of messages. In the current stage, Bi-gram method is used to extract the patterns on two tokens. The future can utilize Tri-gram to interpret 3 tokens patterns. And the multi-word expressions can also be taken into consideration, such as "scrip command" can be interpreted as one word, similarly as "new york" is understand as one city location not "new" and "york". Furthermore, it is possible to utilize RNN (Recurrent Neural Networks) techniques to provide additional advantage of gathering all pre-order vocabulary into consideration for semantic patterns extraction. And also the applying of LSTM (Long Short-Term Memory) method can provide semantic parsing for the task of mapping contents of messages in the domain of code smells.